Assessing the reliability of artificial neural networks

Abstract

The complex problem of assessing the reliability of a neural network is addressed. This is approached by first examining the style in which neural networks fail, and it is concluded that a continuous measure is required. Various factors are identified which will influence the definition of such a reliability measure. For various situations, examples are given of suitable reliability measures for the multilayer perceptron. An assessment strategy for a neural network's reliability is also developed. Two conventional methods are discussed (fault injection and mean-time-before-failure), and certain deficiencies are noted. From this, a more suitable service degradation method is developed. The importance of choosing a reasonable timescale for a simulation environment is also discussed. Examples of each style of simulation method are given for the multilayer perceptron.<>